gender_classifier.m

function [ preds, acc, decvals ] = verifier( folds, images, output )
%VERIFIER Reads images and classifies based on gender
%   Read certain facial features (eyes, nose, mouth, cheek)
%   Train features using regression techniques in order to determine if
%   male or female

vl_setup()

% Read the folds
q = read_lfw_folds(folds);
fin = fopen(folds);
nums = textscan(fin, '%d %d', 1, 'collectoutput', true);
num_men = nums{1}(1);
num_women = nums{1}(2);
total = num_men + num_women;
fclose(fin);

preds = [];
acc = [];
decvals = [];

%{ test on each male and female 
tests = [34, 74, 2, 99, 131, 67, 122, 117, 11, 43];

trainlabels = [];
trainfeats = [];
testlabels = [];
testfeats = [];

% train on each fold except when i==j
for j=1:total

    features = [];
    % 20 pictures of each person
    for k=1:20
        % read images and train on features
        a = strcat(images, '/', q.pairs{(j-1)*20+k});
        im1 = single(rgb2gray(imread(a))) ./ 255;

        % Each column of the 'frames' parameter indicates a point at which to extract a descriptor.  The
        % third row holds the 'scale' of the SIFT descriptor, which for reasons not relevant to our
        % usage is one-twelfth of the side length of the SIFT window.
        siftsize = 32;

        % Here, we extract a SIFT descriptor at point (x, y), using a SIFT window that is 32 pixels on
        % each side.

        % left eye
        x = 65;
        y = 95;
        frames = [x; y; siftsize/12; 0];
        [f1, d1] = vl_sift(im1, 'frames', frames);
        d1 = double(d1);

%{
        right eye
        x = 108;
        y = 95;
        frames = [x; y; siftsize/12; 0];
        [f2, d2] = vl_sift(im1, 'frames', frames);
        d2 = double(d2);
%}

        % nose
        x = 88;
        y = 117;
        frames = [x; y; siftsize/12; 0];
        [f2, d2] = vl_sift(im1, 'frames', frames);
        d2 = double(d2);

        siftsize = 24;
        %mouth
        x = 88;
        y = 149;
        frames = [x; y; siftsize/12; 0];
        [f3, d3] = vl_sift(im1, 'frames', frames);
        d3 = double(d3);

        siftsize = 28;
        %left cheek
        x = 57;
        y = 128;
        frames = [x; y; siftsize/12; 0];
        [f4, d4] = vl_sift(im1, 'frames', frames);
        d4 = double(d4);

%{
        %right cheek
        x = 120;
        y = 128;
        frames = [x; y; siftsize/12; 0];
        [f6, d6] = vl_sift(im1, 'frames', frames);
        d6 = double(d6);
%}

%{
        siftsize = 50;
        %forehead
        x = 86;
        y = 45;
        frames = [x; y; siftsize/12; 0];
        [f7, d7] = vl_sift(im1, 'frames', frames);
        d7 = double(d7);
%}

        feature = [d1; d2; d3; d4]';

        features = [features; feature];
    end

    features = mean(features, 1);

    if any(tests==j)
        testfeats = [testfeats; features];
        if j > num_men
            testlabels = [testlabels; -1];
        else
            testlabels = [testlabels; 1];
        end
    else
        trainfeats = [trainfeats; features];
        labels = [];
        if j > num_men
            labels = -1;
        else
            labels = 1;
        end
        trainlabels = [trainlabels; labels];
    end
end

% From libsvm, there are two functions you will need, which you should call as follows:
% train linearly. Why? because it's the only way it could have finished
% in a reasonable amount of time. 
model = svmtrain(trainlabels, trainfeats, '-t 0');

[preds, acc, decvals] = svmpredict(testlabels, testfeats, model);

% where
% - trainlabels is column vector of training data labels (1 for "same", -1 for "different").
% - trainfeats is your normalized training data feature matrix, with one row per training
%   sample.
% - options is a libsvm option string describing which kernel you want to use and various other
%   options.  Run svmtrain() without arguments to see a description of the options.
% - model is the trained SVM returned.
% - testlabels is a column vector with the labels of the test data.  It's used only to calculate
%   the accuracy of the SVM.
% - testfeats is the normalized test data feature matrix.
% - preds is a column vector holding the predicted class labels (1 or -1) for the test data.
% - acc is the accuracy of the predicted labels.
% - decvals is a column vector with the same size as preds, holding the signed distance from the
%   decision boundary for each test sample.  preds is just determined by the sign of decvals.

% print decision values, names, and actual decisions to output file
fTestOut = fopen(output, 'w');
fin = fopen(folds);
nums = textscan(fin, '%d %d', 1, 'collectoutput', true);
%num_men = nums{1}(1);
%num_women = nums{1}(2);
num_men = 7;
num_women = 3;
i = 1;
xp = sort(tests);

d = textscan(fin, '%s', nums{1}(1) + nums{1}(2), 'collectoutput', true);
for iperfold = 1:num_men
    sub = d{1}{xp(i)};
    fprintf(fTestOut, '%f\t', decvals(i));
    fprintf(fTestOut, '%s\t', sub);
    fprintf(fTestOut, '%d\n', 1);
    i = i + 1;
end

for iperfold = 1:num_women
    sub1 = d{1}{xp(i)};
    fprintf(fTestOut, '%f\t', decvals(i));
    fprintf(fTestOut, '%s\t', sub1);
    fprintf(fTestOut, '%d\n', -1);
    i = i + 1;
end
fclose(fTestOut);
fclose(fin);